Feature Request: Allow sub-agents to use the model specified in frontmatter/task() — add opt-out for the cost-multiplier guard

[Akhilkokani]

Describe the feature or problem you'd like to solve

Allow sub-agents to use the model specified in their .agent.md frontmatter or task() call's model parameter, by adding an opt-out flag for the cost-multiplier guard that currently silently downgrades sub-agent models to the cheapest available model.

Proposed solution

Add a configuration flag (e.g., allowSubagentModelUpgrade: true in ~/.copilot/config.json or --allow-subagent-model-upgrade CLI flag) that lets users opt out of the cost-multiplier guard that silently downgrades sub-agent models.

Currently, when a session runs on a 0× model (e.g., GPT-4.1), every sub-agent model override — whether via model: frontmatter in .agent.md or the model parameter in task() calls — is silently downgraded to the session default. Process logs confirm:

Downgrading subagent model from "claude-sonnet-4.6" (1x) to session model "gpt-4.1" (0x).

This prevents a common and valuable pattern: tiered multi-agent orchestration — using a free model for routine tasks (reads, searches, navigation) while delegating complex work to premium models (Sonnet for code, Opus for architecture). Users who are aware of the cost implications and willing to spend their premium request budget should be able to opt in.

Example prompts or workflows

1. Tiered delegation (primary use case):
   Session on GPT-4.1 (0×). User says "Review the architecture of this service." GPT-4.1 gathers context (free) then delegates to an advisor agent with model: claude-opus-4.6 (3×). Today: advisor runs on GPT-4.1. Expected: advisor runs on Opus.

2. Custom agent model enforcement:
   .agent.md with model: claude-sonnet-4.6 in frontmatter. When invoked as a sub-agent from a free session, the frontmatter model should be honored. Today: silently downgraded.

3. Cost-conscious exploration + quality implementation:
   User navigates codebase with GPT-4.1 (free), then delegates "Write tests for this component" to a specialist agent with model: claude-sonnet-4.6. Today: tests written by GPT-4.1. Expected: tests written by Sonnet.

4. Plugin agents with model requirements:
   A plugin author distributes agents that require specific models for quality (e.g., a code reviewer that needs Opus). Today: the model requirement is ignored. Expected: honored when the user opts in.

5. Budget-aware workflows:
   Combined with /usage, users can monitor spend and make informed decisions about when to invoke premium sub-agents vs. stay on the free tier.

Additional context

Evidence

• Confirmed behavior via process logs: Downgrading subagent model from "claude-sonnet-4.6" (1x) to session model "gpt-4.1" (0x) — tested on v1.0.28
• Workaround exists but loses features: Spawning separate copilot --agent X --model Y -p "..." processes bypasses the guard (each is a top-level session), but loses session integration: shared context, multi-turn refinement, streaming, /tasks, /diff tracking
• Sonnet session workaround is partial: Setting session to Sonnet (1×) avoids the guard for ≤1× sub-agents, but costs 1× for every prompt including simple reads

Related issues

• Model routing, per-agent model selection, and global hooks support #1354 — Model routing, per-agent model selection, and global hooks support
• Question: Does model: in agent frontmatter and task() calls control model selection? #2564 — Does model: in agent frontmatter and task() calls control model selection? (closed, fix shipped in v1.0.22 but cost guard still applies)
• Plugin agents ignore model: frontmatter #2492 — Plugin agents ignore model: frontmatter (closed)

Suggested configuration shapes

// ~/.copilot/config.json
{ "allowSubagentModelUpgrade": true }

# CLI flag
copilot --allow-subagent-model-upgrade

# Per-agent frontmatter (most granular)
---
model: claude-opus-4.6
force-model: true
---